Automated Segmentation and Augmentation of Adolescent Spinal X-ray Images: Dataset Construction and Deep Learning Applications

The high prevalence of spinal diseases poses a significant challenge to global public health and healthcare resources. X-ray imaging, as a non-invasive diagnostic tool, is extensively used for visualizing spinal pathologies. It plays a key role in assessing the structure and function of the spine, aiding clinicians in devising treatment plans and monitoring recovery progress. With the ongoing development of medical imaging technologies, the automated diagnosis and treatment of spinal disorders have garnered increasing attention. However, the current publicly available X-ray-based spinal datasets are relatively small and lack comprehensive annotated data, which limits the effectiveness of deep learning-based segmentation methods. This paper introduces a large-scale dataset of adolescent spinal X-rays, created to address the challenge of automated spinal disease segmentation. The dataset includes 1,175 images: 657 anteroposterior spinal X-rays (303 normal and 354 with scoliosis) and 518 lateral X-rays (160 normal, 101 with wedge deformities, 84 with spondylolisthesis, and 173 with spondylolysis). In total, over 15,000 vertebral annotations are included. To enhance the dataset, data augmentation techniques were applied, expanding the dataset by three times its original size. Testing with various network architectures has shown excellent results. This dataset is expected to promote research in spinal lesion detection and the development of automated diagnostic support systems, contributing to the advancement of deep learning-based clinical decision support tools.